Extracting Entities
Entities are typically nouns that can represent the name of a person, place, or organization. The entities extracted from text supplement the traditional inverted index used in search engines. For example, in "The headquarters of the UNIDO is located in Vienna," an organization entity and a place entity would be extracted in addition to the keywords. It's difficult to form accurate search-engine queries pertaining to an organization in general using keywords alone. A query such as, "Which international organization has its main office in Vienna," will have some irrelevant hits. The process of extracting entities may be time consuming depending on the type of extraction tool. Tools like GATE (gate.ac.uk), for instance, use a gazetteer including lists of organizations and abbreviations along with pattern-matching code to find entities in text. Other tools such as Lingpipe (www.alias-i.com) use a training model based on context to identify the most likely entity type and can be faster.
The use of either model adds to the overhead and increases the response time. Some Q&A systems pretag the text with the associated entities to reduce the response time. This approach has the advantage of speed, but increases the time to index and size of the index. The second method is to extract entities from just the text of the hits returned in response to the translated query. There is less text to process, however, because the extraction occurs while users are waiting for a response—it's usually slower than the first method.
WordNet
Consider these three questions:
- Who is the Queen of Holland?
- What is the highest mountain in the world?
- When did Microsoft buy iView?
The answers to these questions may not contain the same nouns ("Holland" and "mountain") or verb ("buy"). Passages using synonyms such as the "Netherlands," "peak," and "purchase" may be potential answers.
WordNet (wordnet.princeton.edu) is a popular open-source English database containing over 150,000 words from four parts of speech (nouns, verbs, adverbs, and adjectives) with various relationships between words and meanings (synonym sets or synsets). Many Q&A systems use WordNet, even though it was not designed specifically for any particular linguistic process.
The relationship between words is more commonly known than the relationships between synsets. Although words and synsets have distinctive sets of relationships (Figure 2), there is a many-to-many relationship between words and synsets. A single word such as "purchase" is found in four noun synsets and one verb synset. Each of these synsets represent a single sense of the word. Some senses are used more frequently than others. For example, the verb sense of "purchase" is used more often than any of the noun senses.
